OpenAI's open-source speech-to-text model with multilingual support.
Audio & Voice
Whisper is OpenAI's open-source automatic speech recognition (ASR) model — the transcription engine that quietly powers a huge share of the "AI transcript," "AI captions," and "AI meeting notes" products you have used in the last three years. Released in September 2022 under the permissive MIT license and rapidly iterated since (large-v3 landed in late 2023, and the API-only turbo variant followed), Whisper is genuinely a foundational piece of AI infrastructure rather than a consumer product. There is no Whisper app you subscribe to. There is a research paper, a model on Hugging Face, a self-hostable Python package, and — for developers who prefer not to run their own inference — the OpenAI API endpoint at $0.006 per minute of audio. What matters is that Whisper, on the day it shipped, produced open-source transcription quality that was competitive with (and in many languages ahead of) every paid commercial ASR service, and it has only pulled further ahead since.
What Whisper actually does, cleanly stated, is take an audio or video file — in any of dozens of formats, in any of 99 languages — and produce a transcript. Or a translated transcript (source language in, English text out). Or a subtitle file with timestamps. Or a diarized transcript if you pair it with a separate speaker-diarization model. On English, the flagship large-v3 model transcribes accurately enough that reasonable-quality podcast and lecture audio comes out with word-error rates below 5%. On multilingual audio, it holds up in 40+ languages at a quality that beat every commercial service at launch. And because the model is open source and self-hostable, the marginal cost of transcription — once you have the hardware and the setup — is close to zero.
The one-line positioning: Whisper is the speech-recognition engine you either use directly (as a developer, via the OpenAI API or a self-hosted deployment) or that you consume indirectly (via one of the dozens of "AI transcription" products built on top of it). It is not a product for creators looking for a polished "record a meeting, get notes" app — for that, you want Otter, Fireflies, or one of the products built on Whisper. It is the layer underneath.
Whisper's product surface depends on how you access it. Here are the features that matter across the OpenAI API, the open-source model, and the derivative products built on it.
99 languages transcription and translation. Whisper transcribes in 99 languages and can translate any of those languages into English text as part of the transcription. English is best; Spanish, French, German, Italian, Portuguese, Dutch, Russian, Chinese, Japanese, and Korean are excellent; less-resourced languages get progressively noisier but often still usable.
Timestamped output. Word-level and segment-level timestamps for subtitle generation, video captioning, and audio search. The whisper-timestamped and WhisperX projects extend this with forced alignment for higher accuracy.
Multiple model sizes. Tiny (39M params), Base (74M), Small (244M), Medium (769M), Large (1.55B, in versions v1, v2, v3), and Turbo (a 2024 large-v3-turbo variant that is faster with a small quality tradeoff). Choose the size based on the hardware you have and the accuracy you need — Small runs fine on a laptop CPU; Large needs a GPU with 10GB+ of VRAM for practical throughput.
Robust to real-world audio. Whisper was trained on 680,000 hours of multilingual audio scraped from the web, including a lot of noisy, accented, technical-vocabulary content. The result is a model that handles conference-room audio, mobile recordings, thick accents, and technical jargon better than most commercial ASR services designed on clean studio audio.
Open source (MIT license). You can download the weights, run inference locally on your own hardware, and build commercial products on top of Whisper with no royalty or licensing to OpenAI. This is the reason every open-source transcription tool, and a large share of paid ones, use Whisper under the hood.
OpenAI API endpoint. For developers who do not want to run their own inference, OpenAI hosts Whisper at $0.006 per minute of audio through the /audio/transcriptions and /audio/translations endpoints. Same model, no infrastructure. The API model is a fixed version — controllable, predictable, and rate-limited under standard OpenAI API terms.
Turbo model (API only). A 2024 optimization of large-v3 that runs meaningfully faster at similar quality. Available through the OpenAI API at the same price; not weight-released as of writing.
Ecosystem of derivative products. WhisperX (forced alignment plus diarization), Insanely-Fast-Whisper (highly optimized inference), Whisper.cpp (C++ implementation for CPU and mobile), Distil-Whisper (compressed variant), Buzz (open-source desktop app), and dozens of hosted products (Otter, Fireflies, Notta, Sonix, Fathom, tl;dv, and many others) that use Whisper as their engine.
Whisper's pricing has two dimensions: the open-source model (free) and the OpenAI API (per-minute).
Open-source, self-hosted
OpenAI API
| Endpoint | Price | What you get |
|---|---|---|
| /audio/transcriptions | $0.006/minute | Whisper large-v3-turbo transcription in original language |
| /audio/translations | $0.006/minute | Same, translated to English |
| gpt-4o-transcribe | $0.006/minute | Newer multilingual ASR built into the GPT-4o family |
| gpt-4o-mini-transcribe | $0.003/minute | Cheaper, slightly lower-quality newer variant |
$0.006 per minute is $0.36 per hour of audio. A one-hour podcast episode costs less than half a cent to transcribe. A 40-episode back catalog costs less than $2. A meeting-transcription product processing a million hours of audio per month spends $6,000 on Whisper API costs — usually the cheapest line item in the product.
The honest read: Whisper is close to the cheapest way to do high-quality transcription that has ever existed. Even on the OpenAI API, transcription costs are trivial compared to what commercial ASR services (Rev, Sonix, AWS Transcribe, Google Cloud Speech-to-Text) charged before Whisper existed. If you self-host, the marginal cost approaches zero.
Pros
Cons
Developers building transcription features into their products. The OpenAI API at $0.006/minute is the fastest path to "transcribe audio in your app" — a few API calls and you are shipping.
Companies with high-volume transcription needs willing to self-host. For millions of hours of audio, self-hosting Whisper on your own GPU infrastructure is dramatically cheaper than any hosted ASR service.
Researchers, journalists, and academics with recording archives to transcribe. A batch of interviews or lectures runs through Whisper at negligible cost. Buzz or MacWhisper for casual desktop use; scripts for larger batches.
Podcasters producing transcripts and show notes at scale. Feed Whisper your audio, generate a transcript, run it through an LLM for cleanup and section headers, publish. Total cost per episode: cents.
Video creators generating subtitle files (SRT, VTT). Whisper output includes timestamps that convert cleanly to subtitle formats. Suitable for YouTube auto-captions, film festival submissions, and accessibility compliance.
Multilingual content teams doing translation-in-transit. The /audio/translations endpoint takes any of the supported languages and returns English text, useful for teams doing rapid content review across languages.
Accessibility engineers building real-time captioning tools. With WhisperX and streaming inference, you can build near-real-time captioning at a quality level that was infrastructure-only two years ago.
Whisper's competition splits into three lanes: other ASR APIs and models, hosted transcription products built on Whisper or similar engines, and legacy commercial ASR services.
AssemblyAI — commercial ASR API with strong English quality, built-in diarization, summarization, and content moderation. Higher price than Whisper API but bundles more features per request.
Deepgram — commercial ASR API with strong low-latency streaming performance. Popular for real-time transcription products. Comparable quality to Whisper on English; different pricing model.
Rev.ai — legacy transcription service with an API. Higher quality on the paid human-verified tier; the AI-only tier is competitive with Whisper on English.
Google Cloud Speech-to-Text and Azure Speech — the enterprise incumbents. More expensive per minute, deep language coverage, strong enterprise contracts. Whisper generally matches or beats them on quality; they beat Whisper on compliance certifications and SLAs.
AWS Transcribe — AWS's ASR service. Deep AWS ecosystem integration, higher per-minute cost than Whisper API, roughly comparable quality.
Otter.ai, Fireflies.ai, Fathom, tl;dv — hosted meeting-transcription products, many of which use Whisper under the hood. These are the polished-product alternative for teams that want notes and highlights, not a raw transcription API.
MacWhisper and Buzz — desktop applications wrapping Whisper. The right choice for individual users transcribing local files without writing code.
If you are a developer, start with the OpenAI API. Get an API key, send an audio file to /audio/transcriptions, and see the transcript come back. Cost is a fraction of a cent for a test.
If you are self-hosting, install openai-whisper or faster-whisper. Both are Python packages. faster-whisper is meaningfully faster in practice. You will need a GPU with enough VRAM for the model size you pick.
Pick your model size based on hardware and quality needs. Tiny/Base for CPU-only quick tests. Small/Medium on modest GPUs. Large or Turbo on a GPU with 10GB+ VRAM for production quality.
For desktop use with no code, try MacWhisper (macOS) or Buzz (cross-platform). Both wrap Whisper in a polished desktop app for transcribing local audio files.
For diarization (who said what), add WhisperX or a separate speaker-diarization model. Whisper alone gives you a transcript with timestamps; WhisperX adds speaker labels.
For polished, hosted meeting transcription with notes and summaries, use one of the products built on Whisper. Otter, Fireflies, Fathom, and tl;dv all deliver a finished product experience. Whisper is the engine; these are the cars.
Read the OpenAI API docs on file limits. 25MB per request. For longer audio, chunk with a small overlap and stitch results.
Is Whisper free? The model is free and open source under MIT license. Running it costs whatever your compute costs. The OpenAI API is not free — it costs $0.006 per minute of audio.
Can I use Whisper commercially? Yes. MIT license permits commercial use of the model. The OpenAI API is governed by OpenAI's standard API terms, which permit commercial use.
How accurate is Whisper compared to commercial ASR? On English, Whisper large-v3 typically achieves 3-5% word error rate on clean audio and higher on noisy or accented audio. This is competitive with or better than every major commercial ASR service. On less-resourced languages, Whisper often beats commercial services by a wide margin.
Does Whisper support real-time streaming? Not natively. Whisper is designed for file-based inference. Real-time transcription requires chunking the audio stream and passing chunks through Whisper, which introduces latency. For low-latency use cases, Deepgram or a Whisper-streaming implementation (WhisperLive, Faster-Whisper-Server) is a better fit.
Does Whisper do speaker diarization? No, not natively. Whisper produces a single transcript without speaker labels. Pair it with WhisperX or Pyannote for speaker diarization.
What is Whisper Turbo? A 2024 optimization of large-v3 that runs faster with only a small quality decrease. Available on the OpenAI API. Weights were released for the large-v3-turbo variant in late 2024.
Can I fine-tune Whisper? Yes. Whisper is a standard PyTorch model and can be fine-tuned on domain-specific audio using Hugging Face's Transformers library. Common for medical, legal, and technical vocabulary use cases.
What are the biggest Whisper failure modes? Silence hallucination (generating plausible-sounding transcript from silence or music) and word timestamp drift on long audio. Both are mitigated by preprocessing (voice-activity detection to remove silence) and by using WhisperX for forced alignment.
Whisper via the OpenAI API at $0.006/minute is the correct default for any developer building transcription into a product, and self-hosted Whisper is the correct default for any team with a high-volume transcription workload and the engineering capacity to run it. There is no serious price-versus-quality competitor in 2026 for high-quality multilingual transcription. Every commercial ASR service either matches Whisper's quality at higher cost, or falls behind Whisper's quality at similar cost. The market has adjusted; most "AI transcription" products you know are running Whisper under the hood.
Where Whisper stops being the right answer: real-time low-latency transcription at scale (Deepgram is engineered for this), speaker-diarized meeting transcripts (Otter, Fireflies, and Fathom deliver a finished product), teams that need enterprise compliance certifications (Google, AWS, or Azure ASR services have deeper compliance stories), and non-technical users who want a polished desktop or mobile app (MacWhisper, Buzz, or one of the hosted products are the right layer).
The honest recommendation: if you are technical, add the OpenAI /audio/transcriptions endpoint to your stack today and see how cheap transcription has become. If you are not technical and want to transcribe your own audio files, MacWhisper or Buzz gets you there with no code. If you want meeting notes without any setup, use one of the products built on Whisper — Otter, Fireflies, Fathom, or tl;dv. And if you want the highest-quality voice generation to complement Whisper's transcription in a full audio product stack, pair it with ElevenLabs for TTS and — if music is involved — Suno or Udio for scoring.